Wasserstein Bounds for Euler and Randomized Midpoint Langevin Sampling with Noisy Gradients

arXiv:2610.37858 · math.ST, cs.NA, math.NA, stat.CO, stat.TH · Submitted 2026-10-27 · Read on arXiv

math.ST, cs.NA, math.NA, stat.CO, stat.TH

Submitted: 2026-10-27

Updated: 2026-10-27

Comments: 26 pages, 3 figures

License: http://creativecommons.org/licenses/by/4.0/

The gist: Stochastic-gradient Langevin methods are sampling algorithms that replace the gradient of the log probability density by a noisy, possibly biased evaluation.

Abstract

Stochastic-gradient Langevin methods are sampling algorithms that replace the gradient of the log probability density by a noisy, possibly biased evaluation. We study how such gradient-evaluation errors affect the Euler and randomized midpoint discretizations of Langevin dynamics by comparing each noisy chain with its exact-gradient counterpart. For continuously differentiable, strongly log-concave targets with globally Lipschitz gradients, we derive comparison bounds under a sub-Gaussian condition on the gradient noise. The resulting 2-Wasserstein bounds hold uniformly over iterations, and are linear in both the step size and the L squared norm of the evaluation bias, with a square-root dependence on dimensionality. To the best of our knowledge, no comparable 2-Wasserstein bounds have previously been established for randomized midpoint with noisy gradients. For Euler, our results complement recent Wasserstein bounds by allowing biased evaluations, and by not requiring additional regularity assumptions on the evaluation errors.

Related papers